Finding the invisible defects in silicon with deep learning from HAADF-STEM focal series

Silicon is the silent backbone of modern technology, the material that powers everything from the satellites orbiting Earth to the microchips in our phones. Yet, this material is not indestructible. When exposed to the harsh environment of space or the intense radiation of a nuclear reactor, invisible damage begins to accumulate. High-energy particles strike the silicon atoms, knocking them out of their perfect grid-like arrangement and creating tiny, disordered clusters. These clusters are the primary scars of radiation, and they can slowly degrade the performance of the devices they inhabit. The problem for scientists is that these scars are often too faint to see. In a standard electron microscope image, which acts like a powerful camera for the atomic world, these damaged regions produce such weak signals that they blend seamlessly into the background, effectively disappearing from view. Without being able to see these defects, engineers struggle to understand how radiation damages materials or to predict how long a device will last in a hostile environment.

To solve this problem of invisibility, researchers at Aalto University in Finland have developed a new way to look at silicon using a combination of advanced imaging and artificial intelligence. They focused on a specific type of electron microscope technique called high-angle annular dark-field scanning transmission electron microscopy. Imagine taking a series of photographs of a single object, but instead of moving the camera, you slightly change the focus of the lens for each shot. In this method, the microscope takes a stack of images at different focus levels, creating a focal series. While a single image might show nothing but a blurry, uniform gray, the subtle changes in brightness and contrast across the entire stack contain hidden clues about the three-dimensional shape and location of the defects. However, these clues are so faint and buried in noise that the human eye cannot piece them together.

The researchers trained a computer system, a type of deep learning model known as a neural network, to act as a translator for these hidden signals. They did not start with real-world images, but with a massive library of simulated data. Using powerful computer simulations, they modeled how radiation knocks atoms around in silicon, creating realistic clusters of damage. They then simulated what an electron microscope would see if it photographed these simulated clusters, generating thousands of focal series that included the same types of blur and noise found in real experiments. The computer learned to recognize the specific patterns of light and shadow that these invisible defects cast across the different focus levels. Once trained, the system could look at a new stack of images and predict exactly where the defects were, creating a map for each focus plane.

The results of this approach are a significant step forward in seeing the unseen. When the researchers applied their trained system to the simulated data, it successfully reconstructed the three-dimensional shape of the defect clusters. The system could pinpoint the center of a damaged region with an error of less than one nanometer, which is roughly the width of a few silicon atoms. It also estimated the size and shape of the clusters with reasonable accuracy, even when the images were noisy and blurry, mimicking the difficult conditions of a real laboratory. The researchers found that while the system could not perfectly reconstruct every single atom of the defect, it could reliably identify the overall volume and orientation of the damage. This means that instead of seeing nothing, scientists can now generate a coarse but useful 3D model of the damage, telling them not just that a defect exists, but roughly how big it is and where it sits within the material.

This work suggests a new path for characterizing radiation damage that was previously impossible. The method does not rely on finding a single, sharp image of a defect, which often does not exist. Instead, it gathers weak, scattered information from a whole series of images and uses the learned patterns of physics to assemble a coherent picture. The researchers are careful to note that their results come from simulations, and the next step will be to test the system on real silicon samples that have been irradiated. If the method holds up in the real world, it could provide a vital tool for ensuring the reliability of electronics in space and nuclear applications, turning invisible threats into visible, measurable data. By making the invisible visible, this approach offers a way to understand the fundamental limits of the materials that power our technological world.